international association of scientific innovation and...
TRANSCRIPT
International Association of Scientific Innovation and Research (IASIR) (An Association Unifying the Sciences, Engineering, and Applied Research)
International Journal of Emerging Technologies in Computational
and Applied Sciences (IJETCAS)
www.iasir.net
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 11
ISSN (Print): 2279-0047
ISSN (Online): 2279-0055
Error Propagation of Quantitative Analysis Based on Ratio Spectra J. Dubrovkin
Computer Department, Western Galilee College
2421 Acre, Israel
Abstract: Error propagation of the quantitative analysis of binary and ternary mixtures based on the ratio
spectra and the mean-centred ratio spectra has been studied. Gaussian doublets and triplets were used as
models of the mixture pure-component spectra. The mixture spectra were disturbed by random constant and
proportional noises and unknown background. The perturbations of the calibration matrix were modelled by
systematic errors caused by the wavelength shifts. The least-squares estimation of the concentration vector and
the estimation errors were obtained theoretically and numerically. The condition number of the matrix of the
pure-component ratio spectra was theoretically evaluated for binary mixtures. The advantages and
disadvantages of the ratio spectra methods are discussed.
Keywords: quantitative spectrochemical analysis, ratio spectra, mean centred ratio spectra, errors, condition
number, random constant and proportional noise, unknown background.
I. Introduction
One of the main problems of the spectrochemical analysis of white multicomponent mixtures is overlapping of
pure-component spectra which are known a priori. In the period of more than half a century, analysts have
attempted to solve this problem by developing numerous smart mathematical algorithms for processing the
mixture spectrum in conjunction with physical-chemical treatment of the mixture to be analyzed [1]. These
algorithms may be divided into two main groups:
1. Allocation of the analytical points set (or its linear transforms) in the mixture spectrum free of overlapping
with respect to a given analyte.
2. Direct and inverse calibration methods based on solving linear equation systems for data sets of calibration
mixtures with known spectra and concentrations.
The most popular methods of the first group include derivative spectroscopy [2], the method of orthogonal
transforms (“the net analyte signal”) [3, 4], and different modifications of the optical density ratio method [1, 5-
15].
A major progress in developing the methods of the second group was achieved in 1980s due to applying
statistical methods of chemometrics (regularization, principal component analysis, and partial least squares
regression) to combined analytical data obtained by spectroscopic and non-spectroscopic measurements [16].
The success of this approach is attributed to using increased information for analytical purposes. Unfortunately,
in some real-life cases, mixtures which contain different concentrations of pure components are not available
and/or the preparation of artificial mixtures is too complicated. The analysis of medicines with claimed
compositions also requires special approach [17].
In view of the above, many researchers attempted to improve the "old" analytical methods developed in the
“pre-computer era” by using modern instrumentation and computational tools. An interesting practical
applications in this field is the centering modification of the ratio spectra (RS) method [5] (MCRS method [6-
15]). However, its effectiveness was proved only by experimental studies. The goal of our work was giving
rigorous mathematical foundation for studying the noise-filtering properties (error propagation) of the method.
Standard notations of linear algebra are used throughout the paper. Bold upper-case and lower-case letters
denote matrices and vectors, respectively. Upper-case and lower-case italicized letters denote scalars. All
calculations were performed and the plots were built using the MATLAB program.
II. Theory
Consider the spectrum of an additive binary mixture that obeys the Beer-Bouguer-Lambert law:
where is a column vector ( ), and are the column vectors ( ) which represent the
spectra of the first and the second mixture components, respectively, is the concentration vector,
and are the component concentrations, and is the symbol for transpose. The path length is assumed to
be equal to unit. Suppose that the error-free pure-component spectra are a priory known.
Multiplying Eq. 1 by the diagonal matrix
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014,pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 12
where is the -element of , we obtain the ratio spectrum:
where is the unit vector ( ). Constant term can be eliminated by the mean centering of Eq. 3
(subtracting its average):
where chevron is the mean value symbol and Matrix equation (4) consists of linear
equations for each wavelength. Earier it was suggested to evaluate unknown concentration ( at one optimal
wavelength (e.g., at the point of the maximum) by linear calibration procedure using standard mixtures
[6]. Such calibration can significantly reduce the systematic errors [18]. On the other hand, the single-point
analysis results in the losses of information that is contained in the rest of the analytical points. Therefore, we
prefer to solve Eq. 4 by the least squares (LS) method using the “best” combination of the analytical
wavelengths (e.g., [19]).
The LS-solution of Eq. 4 is [20]
where .
Since
where Similarly,
where
and are obtained by changing the indexes in Eqs. 4
and 5.
The above algebraic operations represent, actually, linear transformation of Eq. 1. According to statistical
concepts [20], in the presence of uncorrelated normal noise (perturbation) with zero mean, the LS-estimate is
the best linear unbiased estimate with minimal dispersion, which cannot be decreased by any linear
transformations. However, for other kinds of noise (e.g., proportional), this conclusion is not valid.
The LS-solution of Eq.1 gives two calibration vectors ( :
which differ from the corresponding vecrors obtained by the MCRS method (Eqs. 6 and 7).
If a mixture spectrum contains uncorrelated normal noise with zero mean and constant dispersion , the mean
squared error of the LS estimation of the component concentration vector depends on the sum of the
diagonal elements of inverse matrix (8) [20]:
If the standard deviation of the noise is proportional to the response of the spectral instrument,
where is the coefficient of the noise.
Similarly, if the MCRS method is used,
To compare the LS and the MCRS methods, we used the standard deviation (std) ratio:
The mathematical analysis of the mean centering of ternary mixture ratio spectra is similar to the corresponding
analysis for binary mixtures (see APPENDIX A). In the case of ternary mixtures, the error term of the third
component will appear in Eqs. 9-12.
The error propagation in multicomponent analysis can be studied using the condition number of matrix It is known that the relative prediction uncertainty in quantitative analysis [21] is
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014,pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 13
where
and
are the relative uncertainties of matrix (calibration errors) and of the mixture spectra
(measurement errors), respectively. Since the theoretical calculation of the condition number is possible only in
some simple cases (see APPENDIX B), this number is generally evaluated numerically by computer modeling.
III. Computer modeling
A. Binary mixtures
To perform computer modeling, the components of a symmetrical Gaussian doublet (Fig. 1) were chosen as
elements of matrix :
where is the abscissa of the spectrum plot (e.g., wavelength),
is full width at half maximum of Gaussian lines, and are the positions
of the component maxima.
The condition number of ratio spectra matrix was evaluated (APPENDIX B) as
where is the sampling interval along -axis. Since the value could
not be calculated analytically, the ratio
was calculated numerically (Fig. 2).
From the curves shown in Fig. 2, it can be concluded that the RS method is a little less error-sensitive than the
LS method ( ) only for a large number of analytical points in the case of a strongly overlapping
Gaussian doublet ( ). For a resolved doublet ( ), the RS method is not effective.
The standard deviation ratios (Eq. 13) were evaluated for different sets of analytical points. The chosen sets
were located in the neighborhood of the doublet middle point (Fig. 3) and symmetrically around the point 598
of the doublet MCRS (Fig.1). Since the wings of Gaussian lines quickly decay to zero, the intensity of the
MCRS approaches infinity. To partly compensate for this drawback, very small constant background (0.001)
was added to the doublet components (Eq. 15).
Based on the results presented in Figs. 4 and 5, it can be concluded that it is more advantageous to apply the LS
method to the MCRS than to the original spectra only in the case of proportional noise (which is typical of UV-
VIS instruments). The sets of analytical points were identical for both LS and MCRS methods.
The systematic errors caused by uncompensated second-order polynomial background in the mixture spectrum,
were evaluated numerically for both methods. The obtained values were close to 1.
Another type of errors is caused by the systematic errors in the matrix of the pure-component spectra. While the
random errors of matrix can be significantly decreased by averaging, the systematic errors cannot be
eliminated. The main source of the latter type errors is the shift of the spectral points from their original position
Figure 1. Symmetrical Gaussian doublet with Figure 2. Dependences of relative condition number on
constant base line and its MCRS for different integration limits
( ∙∙∙) Doublet ( , base line = 0.001); and (c) 0.2. ( — ) the MCRS.
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014,pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 14
Figure 3. Analytical points for binary mixtures. Figure 4. Dependences of std ratio on for
different analytical sets
Mixture ( ) and pure-component ( and ) spectra. Constant and proportional noise (above and below the line
Five-point (circles) and nine-point analytical sets respectively). The designations of the analytical sets (rectangles): 1a, 1b, 2a and 2b; a - black, b-red. are the same as in Fig. 3.
Figure 5. Dependences of std ratio on foranalytical sets symmetrically located around point 598
Constant and proportional noise (above and below the
respectively). The set size is given next to each plot.
along -axis This shift depends on the slope of the spectral curve [1]. The theoretical study of the impact of the
wavelength shift in MCRS on the concentration errors is presented in APPENDIX (Eq. C7). It was shown that
the errors of the calculated concentrations strongly depend on the derivatives (slopes) of the mixture spectrum
and of the pure-component spectra. Thus the metrological characteristics of the RS method can be improved by
precise setting of the mixture spectrum and of the pure-component spectra at the same wavelengths. This result
is in agreement with that of the well-known study of the influence of the calibration spectra wavelength shift on
the validity of multivariate calibration models [22, 23]. The impact of wavelength shift on the uncertainty of the
LS-based analysis of spectra and of the corresponding ratio spectra (including MCRS) was also evaluated
numerically. The analytical points were chosen in the range of steep slope of the spectra (Fig. 6, the sampling
interval was halved). The obtained results show that, in some cases, the mean centering procedure significantly
decreases the RS analysis uncertainly (Fig. 6). However, the MCRS method has no advantages over the
common LS method. Moreover, the selection of analytical points is a critical factor of the analysis. For
example, for symmetrical distribution of the analytical points around the point 598 (Fig.1), the LS method gives
better results than the MCLS one, especially for large data sets (Fig. 7).
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014,pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 15
Figure 6. Evaluation of error propagation of binary mixture analysis using different sets of analytical points
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014,pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 16
Left-hand panels: Sets of the analytical points of matrix (∙∙∙) and of the first doublet component RS (—). Ranges of analytical points:
(a) 1090 - 1190, (b) 950 – 1050, (c, d) 850 – 1150. (a-c) and 2 (d). Right-hand panels: (top) and (bottom)
plots.
Figure 7. Dependences of std ratio on for calibration matrix errors
LS (— ) and MCRS (∙∙∙) methods. The analytical sets are located symmetrically around point
598 (Fig.1). The set sizes are 3, 5, 9, and 21 from bottom to top, respectively.
B. Ternary mixtures
The components of two Gaussian triplets were taken as elements of two matrices (Fig. 8, a1 and a2, top
plots). The condition number of the first triplet, is very large due to strong overlapping of the
pure-component spectra. Overlapping in the second triplet is small, consequently, From the
full-region MCRSs (Fig. 8 a1 and a2, bottom plots), “the best” combination of analytical points was selected
empirically (Fig. 8, b1-d1, b2-d2). The points were selected in the regions of the maximum intensity of the
transformed spectra, according to the minimum error criteria for evaluated concentrations. The total prediction
error (Eq. C9) was estimated by averaging over 100 statistically independent numerical experiments using a
95%-confidence interval.
The results presented in Fig. 8 show that the advantage of the MCRS method is significant compression of
spectral data. In other words, this method allows replacing full-range spectra by a relative small number of
analytical points. However, the intensities of the transformed spectra decrease notably, which results in an
increased impact of small errors in MCRS on the of the quantitative analysis uncertainty. It was found that data
compression is achieved at the expence of biased estimation of the second component and large increase of the
total prediction error of the component concentrations for ternary mixtures.
The prediction errors, calculated for 16 ternary mixtures (Table 1), are listed in Table 2. According to these
results, the errors of the LS analysis based on full-range spectra are significantly less than those of the MCRS
method. However, in the same spectral region, the MCRS method is more preferrable than the LS one in the
case of strongly overlapping pure-component spectra.
The most critical factor for the MCRS method are systematic errors of the pure-component spectra matrix. For
example, for the relative errors of the total prediction error are more than 100%.
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014,pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 17
In conclusion, it should be pointed out that the MCRS method employs a very limited number of analytical
points, whose location has been a priory chosen. In contrast to this, in the multivariate regression method, all
possibly relevant spectral and non-spectral data are pooled together. Therefore, generally speaking, the latter
method is preferred to the MCRS analysis. However, consuming large data volumes from analytical instruments
creates a new data-management level. In this connection, traditional spectochemical professionals often prefer
"single-point" analytical methods to complex multi-point procedures.
Figure 8. Gaussian triplet component spectra and their full-region MCRS.
(a1, a2) Component spectra ( top plots) and MCRS ( bottom plots). Analytical point sets for: (b1-d1) a1 triplets and for (b2-d2) a2 triplets, respectively.
Table 1. Mixture concentrations
0.05 0.05 0.05 0.1 0.1 0.8 0.1 0.2 0.7 0.1 0.3 0.6 0.2 0.2 0.6 1/3
0.05 0.9 0.05 0.1 0.8 0.1 0.2 0.1 0.7 0.3 0.1 0.6 0.2 0.6 0.2 1/3
0.9 0.05 0.9 0.8 0.1 0.1 0.7 0.2 0.1 0.6 0.3 0.1 0.6 0.2 0.2 1/3
Table 2. Total prediction errors, %
Disturbance
LS (full)
LS (part)
MCRS
Constant noise )
0.35±0.13 83±30 5.0±1.6
0.027±0.091 0.20±0.071 0.41±0.12
Proportional noise
(
0.25±0.088 67±27 4.3±1.3
0.017±0.0059 0.15±0.054 0.31±0.096
Constant noise ) Background
( )
0.41±0.14 79±29 5.5± 1.7
0.16±0.059 0.25±0.081 0.46±0.14
Wavelength shift
0.019±0.0075 0.057±0.025 0.30±0.12
0.037±0.0144 0.12±0.056 0.61±0.25
0.016±0.0065 0.016±0.0071 0.070±0.018
0.098±0.039 0.28±0.13 64±128
0.040±0.016 0.041±0.017 0.70±1.3
Data for two calibration sets (Fig. 8, a1 and a2) are given in the upper and lower rows, respectively.
References [1] I.Ya. Bernstein and Yu.L. Kaminsky, Spectrophotometric Analysis in Organic Chemistry. Leningrad: Science, 1986.
[2] J. M. Dubrovkin, V. G. Belikov, Derivative Spectroscopy. Theory, Technics, Application. Russia: Rostov University, 1988. [3] J. M. Dubrovkin, “The possibility of the discrete Fourier transform-based quantitative analysis for overlapping absorption
bands”, Izvestia Severo-Kavkazskogo Nauchnogo Centra Vysšey Školy, Yestestvennye Nauki, N1, 1981, pp. 57-60.
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014,pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 18
[4] A. Lorber, “Error propagation and figures of merit for quantification by solving matrix equations”, Anal. Chem., vol. 58, 1986,
pp. 1167–1172 [5] M. J. S. Dewar and D. S. Urch, “Electrophilic substitution. Part VIII. The nitration of dibenzofuran and a new method of
ultraviolet spectrophotometric analysis of mixtures”, J. Chem. Soc., N1, 1957, pp. 345-347.
[6] A. Afkhami and M. Bahram, “Mean centering of ratio kinetic profiles as a novel spectrophotometric method for the simultaneous kinetic analysis of binary mixtures”, Anal. Chim. Acta, vol. 526, 2004, pp. 211-218.
[7] A. Afkhami and M. Bahram, “Mean centering of ratio spectra as a new spectrophotometric method for the analysis of binary and
ternary mixtures”, Talanta, vol. 66, 2005, pp. 712-720. [8] H. M. Lotfy and M. A. Hegazy, “Comparative study of novel spectrophotometric methods manipulating ratio spectra: An
application on pharmaceutical ternary mixture of omeprazole, tinidazole and clarithromycin”, Spectrochim. Acta, A: Molecular
and Biomolecular Spectrosc.,.96, 2012, pp. 259-270. [9] E. A. Abdelaleem and N. S. Abdelwahab, “Simultaneous determination of some antiprotozoal drugs in different combined
dosage forms by mean centering of ratio spectra and multivariate calibration with model updating methods”, Chem. Central J.,
vol. 6:27, 2012, pp.1-8. [10] M. M. Issa, R. M. Nejem, A. M. Abu Shanab and N. T. Shaat,” Resolution of five-component mixture using mean centering
ratio and inverse least squares”, Chem. Central J., vol. 7:152, 2013, pp.1-11.
[11] H. W. Darwish, S. A. Hassan, M. Y. Salem and B. A. El-Zeiny, “Three different spectrophotometric methods manipulating ratio spectra for determination of binary mixture of Amlodipine and Atorvastatin”, Spectrochim. Acta, A: Molecular and
Biomolecular Spectrosc.,83, 2011, pp. 140-148.
[12] N. M. Bhatt, V. D. Chavada, M. Sanyal and P. S. Shrivastav, “Manipulating Ratio Spectra for the Spectrophotometric Analysis of Diclofenac Sodium and Pantoprazole Sodium in Laboratory Mixtures and Tablet Formulation”, The Scientific World Journal,
Vol. 2014, 2014, Article ID 495739, 10 pp.
[13] H. M. Lotfy, N. Y. Hassan, S. M. Elgizawy and S. S. Saleh, “Comparative study of new spectrophotometric methods: an application on pharmaceutical binary mixture of ciprofloxacin hydrochloride and hydrocortisone”, J. Chil. Chem. Soc. vol.58 ,
2013, pp. 1651-1657 .
[14] H. M. Lotfy, M. A.l. M. Hegazy and S. A. N. Abdel‐Gawad, “Simultaneous determination of Simvastatin and Sitagliptin in tablets by new univariate spectrophotometric and multivariate factor based methods”, European J. Chem. vol .4 ,2013, pp. 414-421.
[15] N. W. Ali, M. A. Hegazy, M. Abdelkawy and E. A. Abdelaleem, “Simultaneous determination of methocarbamol and
ibuprofen or diclofenac potassium using mean centering of the ratio spectra method “, Acta Pharm.vol. 62, 2012, pp. 191–200. [16] H. Martens and T. Næs, Multivariate Calibration. New York: Wiley, 1992.
[17] “US Pharmakopia <197> Spectrometric identification tests”, http://www.pharmacopeia.cn/v29240/usp29nf24s0_c197.html
[18] J. Dubrovkin, “Evaluation of the maximum achievable information content in quantitative spectrochemical analysis”, International Journal of Emerging Technologies in Computational and Applied Sciences, vol. 1-6, 2013, pp. 1-7.
[19] J. M. Dubrovkin, “Quantitative analytical spectrometry of multicomponent systems with known quantitative composition using
the orthogonal projection method”, Zhurnal Prikladnoi Spectroscopii, vol. 50, 1989, pp. 861-864. [20] G. A. F.Sever and A. J. Lee,Linear Regression Analysis, 2nded.,New Jersey:John Wiley and Sons, 2003.
[21] K. Danzer, M. Otto, L. A. Currie, “Guidelines for calibration in analytical chemistry. Part 2. Multispecies calibration (IUPAC
Technical report,” Pure and Applied Chemistry, vol. 76, 2004, pp. 1215-1225. [22] H. Swierenga A.P. de Weijer, R.J. van Wijk b and L.M.C. Buydens, “Strategy for constructing robust multivariate calibration
models”, Chem.Intel. Lab. Syst., vol. 49, 1999, pp. 1–17.
[23] F. Vogt and K. Booksh, ” Influence of wavelength-shifted calibration spectra on multivariate calibration models”, Appl. Spectrosc., 2004, vol. 58, pp.624-635.
Appendix
A. Mean centering of ratio spectra for ternary mixtures [6].
Similar to the case of binary mixtures (Eq. 1), consider the spectrum of an additive ternary mixture which obeys
the Beer-Bouguer-Lambert law:
1st step: multiplication of Eq. A1 by matrix (2):
and mean centering of (A2):
2nd
step: multiplication of Eq. A3 by diagonal matrix with non-zero elements
:
where and mean centering of (A4):
where The LS solution of Eq. A5 is
where
and
The slope and the intercept of the linear equation for the first and the second components is readily obtained in
the same way.
B. Condition number of the RS matrix of the pure-components of a Gaussian doublet.
Transformed matrix (Eq. 15) has the form
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014, pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 19
where is defined by Eq. 2. Then
where For sufficienly large the sums in Eq. B2 can be substituded by integrals:
where and The condition number of matrix
is
where λ is the eigenvalue of matrix . The eigenvalues are the solution of the following
equation:
where det is the determinant symbol and is the identical matrix. From Eqs. B2-B6, we obtain:
Using Taylor series, it is easy to show for that
Substituting Eq. B8 into Eq. B7, we obtain:
For approximation B9 is very close to the precise value obtained numerically.
Due to the symmetry of the Gauss doublet the same result was obtained for
C. Gaussian doublet. Case study:Impact of systematic errors of matrix S on the errors of binary mixture
analysis
Let
where is disturbance of matrix . The elements of matrix can be regarded as known constants Substituting Eq. C1 into Eq. 1, we obtain:
where is an unknown concentration vector. The LS solution of Eq. C2 is
Suppose that the shift of point along -axis from its original position in the mixture spectrum is the main
source of systematic error . This shift, being dependent on the slope of the spectral curve [1], is measured by
the fraction of sampling interval along -axis:
where k is a constant, is the derivative of the spectrum in the point .
Setting Eqs. C1 and C4 into Eq. C3 and using the matrix equation for small , we obtain:
Next, neglecting the small terms that contain of the order , we have
where .
Eq. C6 is identical to the equation obtained in the case of the round–off errors of the explanatory variables (the
elements of the regression matrix) [20].
Setting Eq. B1 in , we have for the ratio method:
where
J. Dubrovkin, International Journal of Emerging Technologies in Computational and Applied Sciences, 9(1), June-August, 2014, pp. 11-20
IJETCAS 14-507; © 2014, IJETCAS All Rights Reserved Page 20
The first term in the brackets in Eq. C7 appears due to the shifts of the mixture pure-component spectra.The
second term is connected with the errors of the mixture ratio spectra that are due to the shift of the second pure-
component spectrum.
To compare errors and
we used the ratio of the vector norms
which was evaluated numerically.
Besides, total prediction error for mixtures was calculated as
where and